arXiv:1604.08709v3 [cs.AI] 21 Nov 2016 etns uha itiue ytm n mefc inform imperfect and c systems its [ distributed and as knowledge capture such to settings, fields various to applied fully htagent that oedtie ro fProposition of proof detailed more a epu omnso nerirvrino hspaper. this of version earlier t an authors on The comments 12&ZD119. helpful projects key NSSF and Professionals rtodrmdllogic modal first-order l seatyteoesuidb agadFan and Wang by studied one the exactly is els eatc a egetysmlfidb nrdcn ternary a that introducing out by turns simplified it greatly be perspective, ne can this the semantics treating Under by diamond. logic modal special normal a disguised a as viewed be Keywords: problems. under hard our previously some sharpens simplifies which and logi logic system, resulting value” modal The normal logic. transparent the whic a of diamond, power value” expressive the “knowing crease the of generalization binary n h nwrt nebdigqeto,sc s“ know “I as such question, embedding an to answer the ing proposed logic value” “knowing conditional logi the modal for that rules show epis reasoning we and nonstandard lo axioms These standard in on. the so interest i.e., and how”, growing “knowing what”, a “knowing whether”, witness years Recent Abstract ⋆ ope modal a s using mainly by Hintikka knowledge and propositional Wright about von patterns la à logic epistemic Classic Introduction 1 ⋆⋆ oes hc soitsoewrdwt two constant with of value world one associates which models, oo.Rcn er ins rwn neeti h oiso logics the in interest growing a witness years Recent on. so know “I true”, 3 ulse ntepoednso iL21 p 6-8 yColl by 362-381 pp. 2016 AiML of proceedings the in Published ajn agakolde h upr rmteNtoa Pro National the from support the acknowledges Wang Yanjing , 9 Koigvle oi sanra oa logic modal normal a as logic value” “Knowing ] .Hwvr neeya ie nweg sotnepesdi expressed often is knowledge life, everyday in However, ). nwn au,nra oa oi,trayrlto,binar relation, ternary logic, modal normal value, knowing nw that knows i what c ne nutv osrit,temdllgcbsdo such on based logic modal the constraints, intuitive Under . orpswr s,“ know “I is”, password your eateto hlspy eigUniversity Peking Philosophy, of Department rpsto stu.Eitmclgchsbe success- been has logic Epistemic true. is proposition a a uadYnigWang Yanjing and Gu Tao 3.5 . i acsil olsta ontareo the on agree not do that worlds -accessible [ 12 , 13 ] how akteaoyosrveesfrtheir for reviewers anonymous the hank oevr hr savr natural very a is there Moreover, . rmfrSeilSpoto Eminent of Support Special for gram g ulctos hsi rf with draft a is This Publications. ege ihtebnr imn has diamond binary the with c a eivld nti paper, this In invalid. be may c opoeteterm and theorem” the prove to relation h rgnlfis-re Kripke first-order original the ,srrsnl,de o in- not does surprisingly, h, aino h voeao as operator Kv the of gation isaeuulyntnormal, not usually are gics yWn n Fan and Wang by ei oiso “knowing of logics temic tnigo h “knowing the of standing to ae c.e.g. (cf. games ation ag nmulti-agent in hange whether uisteinference the tudies R rator modality, y c i uhknowledge such f em fknow- of terms n nsadr Kripke standard in K h li is claim the i rpemod- Kripke oexpress to [ 12 ] can 2 “Knowing value” logic as a normal modal logic expressions [7,8,12,13,4,5,10]. The fundamental idea is to simply treat “knowing whether”, “knowing what”, “knowing how” as new modalities, just as “knowing that” in standard epistemic logic (cf. the survey [11]). The resulting logics are usually not normal in the technical sense that usual modal axioms and rules may be invalid. For example, the K axiom for normal modal logic is not valid for the knowing whether operator, i.e., Kw(p → q) ∧ Kwp → Kwq does not hold, e.g., knowing that p is false makes sure that you know whether p and also whether p → q, but it does not tell you anything about the truth value of q. Similarly, knowing how to swim and knowing how to cook does not mean knowing how to swim and cook at the same time, thus invalidating Khp ∧ Khq → Kh(p ∧ q), a theorem in normal modal logic when taking Kh as a box modality. On the other hand, the non-normality does not necessarily mean that we have to abandon Kripke models for more general models. As demonstrated in [5], we can still use Kripke models to accommodate those non-normal modal logics by using nonstandard yet intuitive truth conditions for the new modalities. However, there is usually a clear asymmetry between the relatively simple modal language and the “rich” model which may cause troubles in axiomatizing the logic. For example, the conditional knowing value logic proposed in [12] has the following language ELKvr (where i ∈ I, p ∈ P, c ∈ C and I, P, C are countably infinite):

φ ::= ⊤ | p |¬φ | (φ ∧ φ) | Ki φ | Kvi(φ, c) where Kvi(φ, c) says that i knows [what] the value of c [is], given φ, e.g., I know the password of this website given it is 4-digit, since I may have only one 4-digit password ever, although I am not sure which password I used for this website without the information on the digits. The language is interpreted on first-order

Kripke models M = 〈S, D, {→i : i ∈ I}, V, VC〉 where 〈S, {→i : i ∈ I}, V 〉 is a standard Kripke model, and D is a constant domain, and VC assigns to each (non-rigid) c ∈ C an element in D on each s ∈ S. The semantics for the new Kvi operator is as follows:

M , s  Kvi(φ, c) ⇐⇒ for any t1, t2 : if s →i t1, s →i t2, M , t1  φ and M , t2  φ, then VC(c, t1)= VC(c, t2).

According to this semantics, the formula Kvi(φ, c) can also be understood as a first- 1 r order modal formula: ∃xKi(φ → c = x). Thus ELKv can be viewed as a (small) fragment of first-order modal logic where a quantifier is packed with a modality. It is shown in [12] that ELKvr is equally expressive as public announcement logic extended with unconditional Kvi operators proposed in [7] (i.e., only Kvi (⊤, c) r are allowed). Satisfiability of ELKv over arbitrary models is PSPACE-complete, as proved in [2]. Note that although values are assigned to the constants in the model, we cannot talk about them explicitly in the language. In fact, we only care about whether on some worlds a given constant has exactly the same value. The

1 Note that there is a constant domain D and each c has a unique value on each state. 3 contrast between the rich model and the simple language made the completeness proof of the following axiomatization SELKVr S5 quite involved over multi-agent S5 models (cf. [13]). 2 System SELKVr S5 Axiom Schemas Rules TAUT p, p → q all the instances of tautologies MP DISTK Ki(p → q) → (Ki p → Ki q) q T φ Ki p → p NECK 4 Ki p → KiKi p Kiφ 5 φ ¬Ki p → Ki¬Ki p SUB DISTKvr Ki (p → q) → (Kvi (q, c) → Kvi(p, c)) φ[p/ψ] Kvr ψ ↔ χ 4 Kvi (p, c) → KiKvi(p, c) RE Kvr ⊥ Kvi (⊥, c) φ ↔ φ[ψ/χ] Kvr ˆ ∨ Ki(p ∧ q) ∧ Kvi (p, c) ∧ Kvi (q, c) → Kvi (p ∨ q, c)

Since the Kvi operator is not a modality taking only propositions as arguments, it is hard to say whether the above logic is normal or not. DISTKvr looks a little K bit like the axiom but it is in fact about the interaction between Ki and Kvi. Kvr 4 is a variation of the positive introspection axiom, and the corresponding Kvr negative introspection is derivable. ⊥ says that the Kvi operator is essentially Kvr ˆ a conditional. Axiom ∨ handles the composition of the conditions, where Ki is the dual of Ki. In this paper, we look at ELKvr from a new yet “normal modal logic” perspect- ive in order to answer the following questions: (i) Since we do not talk about values in the language, is there a simpler value- free Kripke-model based semantics for ELKvr that can keep the logic (valid formulas) the same? If so, we can restore the symmetry between the lan- guage and the model and understand the essence of our logic. (ii) Can ELKvr be linked to a normal modal logic (modulo some syntactic trans- formation)? If so, we can apply many standard modal logic techniques to simplify previously complicated discussions. We give positive answers to both questions, inspired by a crucial observation: c Observation ¬Kvi(φ, c) can be viewed as a diamond operator ◊i φ which says that there are two i-accessible φ-worlds, which do not agree on the value of c.

Note that to simplify the technical discussion in order to reveal the crucial points, in this paper we focus on the logic over arbitrary models. Our techniques can be applied to the S5 setting. The contributions of this paper are summarized as below: • We give a simple alternative Kripke semantics to ELKvr without value assign-

2 φ[ψ/χ] in the rule RE (replacement of equivalents) denotes any formula obtained by replacing some occurrences of ψ by χ. 4 “Knowing value” logic as a normal modal logic

ments, which does not change the set of valid formulas. The completeness proof is much simpler compared to the one in [13]. • c c We generalize ◊i φ in a natural way to a binary diamond operator ◊i (φ, ψ). It turns out the generalization does not increase the expressive power of the logic but it can give us a transparent normal modal logic proof system. • The normal modal logic perspective helps us to discover a bisimulation notion for ELKvr and obtain a proof system for a weaker language proposed by [7]. Our findings show that ELKvr is essentially a “disguised” normal logic, and this may help us to understand such nonstandard epistemic operators better. The rest of the paper is organized as follows: we first introduce in Section 2 the c language with the unary diamond ◊i and a semantics based on Kripke model with both binary and ternary relations under three intuitive constraints. We show that this semantics is equivalent to the original FO Kripke semantics of ELKvr modulo validity (under a straightforward syntactic translation). In Section 3 we prove the completeness of the translated SELKVr system w.r.t. the new semantics directly. This demonstrates the advantages of using this simplified semantics. In Section 4 c we generalize the ◊i naturally to a binary one and show that the extended lan- guage is in fact equally expressive as ELKvr . On the other hand, the extended language facilitates a transparent normal logic proof system. It then helps us in Section 5 to come up with a notion of bisimulation for ELKvr and obtain a proof system for a weaker language proposed earlier. We conclude in the end with future directions.

2 Negation of Kvi as a diamond

As we mentioned in the introduction, ¬Kvi(φ, c) can be viewed as a diamond c formula ◊i φ: there are two i-accessible φ-worlds which do not agree on the value c c of c. Then ƒi φ := ¬◊i ¬φ means that all the i-accessible ¬φ-worlds agree on the value of c (and it is Kvi(¬φ, c) essentially.). For uniformity of the language, we c r take ƒi as the primitive symbol and introduce the following language (MLKv ):

c φ ::= ⊤ | p |¬φ | (φ ∧ φ) | ƒiφ | ƒi φ where p ∈ P, c ∈ C, i ∈ I. We can, without difficulty, inductively define a translation function T from the original ELKvr to this modal language (the other way is also straightforward):

Definition 2.1 A translation function T from ELKvr to MLKvr formulas is defined as follows: T(p) = p T(¬φ) = ¬T(φ) T(φ ∧ ψ) = T(φ) ∧ T(ψ) T(Kiφ) = ƒi T(φ) c T(Kvi(φ, c)) = ƒi ¬T(φ) 5

Now we have the following translated axioms (the names are kept):

DISTKvr c c T( ) = ƒi(p → q) → (ƒi ¬q → ƒi ¬p) Kvr c c c T( ∨) = ◊i(p ∧ q) ∧ ƒi ¬p ∧ ƒi ¬q → ƒi ¬(p ∨ q) Kvr c T( ⊥) = ƒi (¬⊥)

We can massage the axioms, 3 and obtain the following equivalent system (mod- ulo translation T) from SELKVr (i.e., SELKVr -S5 without the S5 related axioms T, 4, 5, Kvr 4): Rules → MP φ, φ ψ SMLKVr ψ System φ Axiom Schemas NECK TAUT ƒiφ all the instances of tautologies φ DISTK ƒ ƒ ƒ NECKvr i(p → q) → ( i p → iq) c DISTKvr c c ƒi φ ƒi(p → q) → (ƒi p → ƒi q) Kvr c c c φ ∨ ◊i(p ∧ q) ∧ ◊ (p ∨ q) → (◊ p ∨ ◊ q) SUB i i i φ[p/ψ] RE ψ ↔ χ φ ↔ φ[ψ/χ] Note that instead of Kvr ⊥, we have a more classic-looking rule NECKvr . In fact, it is equivalent to have either Kvr ⊥ or NECKvr in the system: from NECKvr , it is trivial to derive Kvr ⊥, and from Kvr ⊥, DISTKvr and NECK it is also straight- forward to derive each instance of NECKvr by taking p in DISTKvr as ⊤. To a modal logician, SMLKVr may look much more friendly compared to r Kvr c SELKV . In particular, ∨ is simply a conditional distribution axiom for ◊i over c c c disjunction. Note that ◊i(p∨q) → (◊i p∨◊i q) is valid but ◊i (p∨q) → (◊i p∨◊i q) is not, e.g., all the p worlds agree on the value of c and all the q worlds agree on the c value of c but they just cannot agree with each other. This demonstrates that ◊i is apparently not a normal modality. However, as we will discover later that this apparent non-normality is a bit misleading and we will restore the normality in c the next section by considering a natural binary generalization of the ◊i operator. Now we are going to give a simplified but equivalent semantics to MLKvr such that the system SMLKVr is sound and complete. The idea is to abandon the first-order Kripke model and use a rather standard Kripke model for propositional modal logics since much of the information in the FO Kripke model is not relevant for the language MLKvr . r c Definition 2.2 A model for MLKv is a tuple 〈S, {→i : i ∈ I}, {Ri : i ∈ I, c ∈ C}, V 〉, where

• 〈S, {→i: i ∈ I}, V 〉 is a standard Kripke model with binary relations. • c For each c ∈ C,Ri is a triple relation over S satisfying for any s, t, u, v ∈ S:

3 r ƒ ƒc ƒc For example, T(DISTKv ) is equivalent to i(¬q → ¬p) → ( i ¬q → i ¬p) under RE, which is equivalent to ƒi (p → q) → (ƒi p → ƒiq) under SUB. 6 “Knowing value” logic as a normal modal logic

c c (i) SYM: sRi tu ⇐⇒ sRi ut c (ii) INCL: sRi tu only if s →i t and s →i u c c c (iii) ATEUC: sRi tu and s →i v imply that at least one of sRi tv and sRi uv holds c Intuitively, sRi tu roughly means that s can see two i-accessible worlds t, u which do not agree on the value of c, although we do not have value assignments for c in the model. Further conditions are imposed to let the ternary relation really capture what we want. (i) is a symmetry condition on the later two arguments c of Ri . Condition (ii) establishes the connection between the ternary and binary relations. The most crucial condition is (iii), an anti-euclidean property 4 that says if two i-accessible worlds do not agree on the value of c then for any third i-accessible world it must disagree with one of the two worlds on c. 5

❢❢❢2 t ❢❢❢2 t ❢❢❢2 t ❢i❢❢❢ ❢i❢❢❢ ④ ❢i❢❢❢ ❢❢❢❢❢ ❢❢❢❢❢ ④④ ❢❢❢❢❢ s ❨❨❨❨❨❨ c s ❨❨❨❨❨❨ ④ c s ❨❨❨❨❨❨ c i❨❨❨❨❨❨, implies i❨④❨④❨❨❨❨, or i❨❨❨❨❨❨, , u c④ , u , u ④ ♥♥♥ i i ④④ i ♥♥♥ ④④ ♥♥c   ④④  ♥♥♥ v v v ♥

c The new semantics is defined as follows which reflects the intuition behind Ri :

M , s  ⊤ always M , s  p ⇔ s ∈ V (p) M , s  ¬φ ⇔ M , s 6 φ M , s  φ ∧ ψ ⇔ M , s  φ and M , s  ψ

M , s ◊iφ ⇔ there exists t such that s →i t and M , t  φ c c M , s ◊i φ ⇔ there exist u, v such that sRi uv, M , u  φ and M , v  φ In the rest of this section, we show that the above semantics of MLKvr is equivalent to the semantics for ELKvr modulo the syntactic translation T. To show this, the difficult part is to saturate an MLKvr model with value assignments while keeping the truth values of formulas modulo translation T. Note that this r c is not straightforward, as it is possible in an MLKv model that sRi tt and there is no way to assign to c two different values on the same world t. Moreover, it can c c happen that sRi uv, s →j u and s →j v while it is not the case that sR j uv. However, we can avoid such problem by preprocessing the MLKvr model before assigning valuations. Lemma 2.3 For any set of ELKvr formula Σ ∪{φ}, Σ  φ iff T(Σ)  T(φ). Proof It suffices to prove that for any set of ELKvr formula Σ, Σ is -satisfiable iff T(Σ) is -satisfiable. We say that an ELKvr model M , s is equivalent to an MLKvr

4 Euclidean property says that ∀x, y, z : xRy ∧ xRz → yRz. Taking R′ = R¯ we have ∀x, y, z : ¬xR′ y ∧ ¬xR′z →¬yR′z, i.e., ∀x, y, z : yR′z → xR′ y ∨ xR′z. Our condition is inspired by this observation. 5 Careful readers may wonder about whether we can break the the ternary relation into two: the i-relation and an anti-equivalence relation. We will come back to this point at the end of the paper. 7 model N , t if for all φ ∈ ELKvr : M , s  φ ⇐⇒ N , t  T(φ). In the following we show that for any pointed FO Kripke model M , s for ELKvr , there is an equivalent MLKvr model N , t, and vice versa. ⇒: For any pointed FO Kripke model M , s, we can naturally define the ternary c c relation Ri as follows: sRi tu iff s →i t, s →i u and VC (c, t) 6= VC (c, u). It’s straight- forward to check that the resulting model N is an MLKvr model (satisfying the three conditions) and N , s is equivalent to M , s. ⇐: Recall that what we need to show is the following: given an MLKvr model c r N = 〈S, {→i: i ∈ I}, {Ri : i ∈ I, c ∈ C}, V 〉 and t ∈ S, find an ELKv model M , s such that M , s  φ ⇐⇒ N , t  T(φ). As mentioned, we need to preprocess N before assigning values. The preprocessing consists of two steps: splitting and unraveling. We first split c the states in N into two copies in order to handle the uRi vv problem mentioned ′ ′ c ′ before. Let N be 〈S ×{0, 1}, {→i : i ∈ I}, {Pi : i ∈ I, c ∈ C}, V 〉, where: • ′ (u, x) →i (v, y) ⇐⇒ u →i v • c c (u, x)Pi (v, y)(w, z) ⇐⇒ uRi vw and (v, y) 6=(w, z) • V ′((u, x)) = V (u) It can be verified that N ′ has the three properties of MLKvr models, 6 and c there is no state v such that uPi vv for any u, v. We can prove the following claim by a simple induction on the structure of MLKvr formulas:

′ N , u ≡MLKvr N , (u, x) where x ∈{0, 1}

c c The only non-trivial case is when φ = ◊i ψ. Suppose N , u  ◊i ψ, then ′ ′ c ′ there exist v, v ∈ S (v and v are not necessarily different) such that uRi vv , ′ c N , v  ψ and N , v  ψ. By the definition of Pi and the induction hypothesis, c ′ ′ ′ ′ ′ c (u, x)Pi (v, 0)(v , 1), N , (v, 0)  ψ and N , (v , 1)  ψ, so N , (u, x) ◊i ψ. Sup- ′ c ′ c ′ pose N , (u, x) ◊i ψ, then there exist (v, y)(v , z) such that (u, x)Pi (v, y)(v , z), ′ c N , (v, y)  ψ and N , (v , z)  ψ. According to the definition of Pi , this en- ′ c ′ tails (v, y) 6= (v , z) and uRi vv . By induction hypothesis, N , (v, y)  ψ and ′  ◊c N , (v , z) ψ, so N , u i ψ. As a simple consequence, ′ N , s ≡MLKvr N , (s, 0)(1) To simplify notation, we shall write (s, 0) as s′ in the rest of this proof. ′ ′ ′ c ′ Now we unravel N at s into M = 〈W, {,→i: i ∈ I}, {Qi : i ∈ I, c ∈ C}, U 〉: i i • ′ ′ 1 k ′ W = {〈s , i1, v1,..., ik, vk〉 : there is a path s → v1 . . . → vk in N }. Note that the trivial path 〈s′〉∈ W, • ′ ′ ′ ′ 〈s , i1,..., vk〉 ,→i 〈s , j1,..., um〉 iff m = k+1, 〈s , i1,..., vk〉 = 〈s , j1,..., uk〉, jm = ′ ′ i and vk →i um in N ,

6 c ′ ′ ′ Take the anti-euclidean property as an example. Suppose (u, x)Pi (v, y)(v , y ) and (u, x) →i (w, z). ′ ′ c ′ ′ c If (w, z) is one of (v, y) and (v , y ), done. If not, then (u, x)Pi (v, y)(v , y ) implies uRi vw, and ′ (u, x) →i (w, z) implies u →i w. By the anti-euclidean property of the original model N , we have c c ′ c c ′ ′ either uRi vw or uRi v w. So either (u, x)Ri (v, y)(w, z) or (u, x)Ri (v , y )(w, z). 8 “Knowing value” logic as a normal modal logic

• ′ c ′ ′ c ′ 〈s , i1,..., vk〉Qi 〈s , j1,..., um〉〈s , l1,..., ln〉 iff vk Pi umln, 〈s , i1,..., vk〉 ,→i ′ ′ ′ 〈s , j1,..., um〉 and 〈s , i1,..., vk〉 ,→i 〈s , l1,..., ln〉, • ′ ′ ′ U (〈s , i1,..., u〉)= V (u). Intuitively, the new model M ′ starts from 〈s′〉, and each state corresponds to a path which is accessible from s′ in N ′. It is not hard to verify the three properties of MLKvr models. 7 By definition, the ,→ skeleton of M ′ is a tree-like structure: acyclic, every state except the root 〈s′〉 can be reached eventually by 〈s′〉 and has one and only one predecessor. It follows that for any u, v in M ′, it is not the case that u ,→i v and u ,→j v for any i 6= j. Now we can prove the following by induction on the structure of φ ∈ MLKvr : ′ ′ ′ M , 〈s , i1, v1,..., ik, vk〉≡MLKvr N , vk In particular,

′ ′ ′ ′ N , s ≡MLKvr M , 〈s 〉 (2)

Now the only thing left is to transform the MLKvr model M ′ into an equivalent ELKvr model M . Basically, we just need to give values to c ∈ C on each state c according to the ternary relations Q j. Let M = 〈W, D, {,→i: i ∈ I}, U, VC〉 where: • ′ W and {,→i: i ∈ I} are exactly the same as in M ; • U = U′; • VC(c, w)= |(c, w)|∼. That is, VC(c, w) is the equivalence class under the equival- ence relation ∼ over C × W defined as: c ∼ = {〈(c, u), (e, v)〉 : c = e, ∃s∃j : s ,→j u, s ,→j v, ∀w ∈ W : ¬wQ juv}∪ {〈(c, u), (c, u)〉 | (c, u) ∈ C × W} • D = {|(c, w)|∼ | (c, w) ∈ C × W}; To make sure M is well-defined, we need to show that ∼ is an equivalence relation. Reflexivity and symmetry are obvious, and for transitivity: If (c, w) ∼

(d, u), (d, u) ∼ (e, v), then c = d = e, there exist s, i such that s ,→i w, s ,→i u c ′ ′ ′ while for any t not tQi wu, and there exist s , j such that s ,→j u, s ,→j v while c ′ for any t not tQ juv. Since every state in W has at most one predecessor, s = s . Since there is at most one relation between two different states, i = j. Therefore s ,→i w, s ,→i v and s ,→i u. Suppose towards contradiction that there exists c c c o ∈ W such that oQi wv, then o = s. Thus sQi wu or sQi uv by anti-euclidean property, contradiction. Therefore (c, w) ∼ (e, v). We still need to verify that this assignment is good, in the sense that: for any ELKvr formula φ, M ′, w  T(φ) ⇐⇒ M , w  φ for any w ∈ W. We prove this by induction on φ and only show the non-trivial case: c If φ = Kvi(ψ, c), then T(φ)= ƒi ¬T(ψ). ′ ⇒: Suppose M , w 6 Kvi(ψ, c) then there exist t, t such that w ,→i t, w ,→i t, M , t  ψ, M , t′  ψ and (c, t) 6∼ (c, t′). According to the definition of ∼, this

7 c ′ Again, take the anti-euclidean property as an example. Suppose uQi vv , u ,→i t. Suppose u = ′ ′ ′ ′ ′ ′ c ′ ′ 〈s ,..., uk〉, v = 〈s ,..., vm〉, v = 〈s ,..., vn〉, w = 〈s ,..., wl 〉. Then uk Pi vm vn and uk →i wl , which c c ′ ′ implies at least one of uk Pi vm wl and uk Pi vnwl holds. This together with u ,→i v, u ,→i v and u ,→i w c c ′ imply either uQi vw or uQi v w. 9

c ′ implies ∃u such that uQi tt . But we have shown that every state has exactly one c ′ ′ predecessor, so u = w, and wQi tt . By induction hypothesis, M , t  T(ψ) and ′ ′  ′ ◊c ′ ƒc M , t T(ψ). Therefore, M , w i T(ψ), i.e., M , w 6 i ¬T(ψ). ′ c ′ c ⇐: Suppose M , w 6 ƒi ¬T(ψ), i.e. M , w  ◊i T(ψ). Then there exist ′ c ′ ′ ′ ′ ′ t, t ∈ W such that wQi tt , M , t  T(ψ) and M , t  T(ψ). So w ,→i t, w ,→i t ′ ′ but (c, t) 6∼ (c, t ), i.e. VC(c, t) 6= VC(c, t ). By induction hypothesis, M , t  ψ and ′ M , t  ψ. Therefore, M , w 6 Kvi (ψ, c). It follows that for any ELKvr formula φ,

M ′, 〈s′〉  T(φ) ⇐⇒ M , 〈s′〉  φ (3)

With (1), (2) and (3), we can now conclude that for any MLKvr model N , t there is always an equivalent ELKvr model M , s and this concludes the proof. ƒ Remark 2.4 The above lemma implies that for any ELKvr formula φ:

 φ ⇐⇒  T(φ) which asserts the validities are the same modulo the translation. We need the stronger version to handle strong completeness later.

3 Completeness of SMLKVr In this section, we show a direct proof of the strong completeness of SMLKVr proposed in the previous section. As we will see, this proof is much simpler com- pared to the original completeness proof of SELKVr in [13] due to the fact that we do not need to construct a FO canonical Kripke model with value assignments anymore. Definition 3.1 The canonical model of SMLKVr is a tuple

c M = 〈S, {→i: i ∈ I}, {Ri : i ∈ I, c ∈ C}, V〉 where: • S is the set of all maximal SMLKVr -consistent sets of MLKvr formulas, • s →i t ⇐⇒ {φ : ƒiφ ∈ s}⊆ t, • c ƒ ƒc sRi tu ⇐⇒ (1){φ : iφ ∈ s}⊆ t ∩ u and (2){ψ : i ψ ∈ s}⊆ t ∪ u, • V(s)= {p : p ∈ s}. c Note that condition (2) for Ri says that if s can see two i-accessible worlds which do not agree on c then at least one should satisfy ψ for each Kvi (¬ψ, c) ∈ s.

Proposition 3.2 The canonical model M is an MLKvr model. c Proof We only need to check the three conditions of Ri . c c (i) sRi uv ⇒ sRi vu: Obvious. c c (ii) sRi uv ⇒ s →i u: By condition (1) in the definition of Ri . 10 “Knowing value” logic as a normal modal logic

c c c (iii) sRi uv and s →i t ⇒ either sRi ut or sRi t v: Suppose not. Then according c c c to the definition of Ri , we have {ψ : ƒi ψ ∈ s} 6⊆ u ∪ t and {ψ : ƒi ψ ∈ ƒc s} 6⊆ v ∪ t. So there exist ψ1, ψ2 ∈{ψ : i ψ ∈ s} such that ψ1 6∈ u ∪ t and ψ2 6∈ v ∪ t. According to the property of maximal consistent sets, this entails ψ1 ∧ ψ2 6∈ u ∪ t and ψ1 ∧ ψ2 6∈ v ∪ t. Now, we distinguish two situations: ◊i(¬ψ1 ∧¬ψ2) ∈ s and ◊i(¬ψ1 ∧¬ψ2) 6∈ s, and go on to show that in both cases we would arrive at contradiction. c Suppose ◊i(¬ψ1 ∧¬ψ2) ∈ s. Note that since ψ1, ψ2 ∈ {ψ : ƒi ψ ∈ s}, c c Kvr c ƒi ψ1 ∈ s and ƒi ψ2 ∈ s. Then according to ∨, we have ƒi (ψ1 ∧ ψ2) ∈ s. c c So ψ1 ∧ ψ2 ∈{ψ : ƒi ψ ∈ s}. Since {ψ : ƒi ψ ∈ s}⊆ u ∪ v, ψ1 ∧ ψ2 ∈ u ∪ v. But this means that ψ1 ∧ ψ2 ∈ u ∪ t or v ∪ t, contradiction. Suppose ◊i(¬ψ1 ∧¬ψ2) 6∈ s, then ƒi(ψ1 ∨ ψ2) ∈ s. According to the c definition of Ri , we have ψ1 ∨ ψ2 ∈ t. By the property of MCS, at least one of ψ1 and ψ2 is in t. However, since ψ1 6∈ u ∪ t and ψ2 6∈ v ∪ t, we have ψ1, ψ2 6∈ t, contradiction. Therefore, the canonical model M is indeed an MLKvr model. ƒ By a Lindenbaum-like argument, every consistent set of MLKvr formulas can be extended to a maximal consistent set (of MLKvr formulas). In the following c we (as routine) prove the existence lemma for both modalities ◊i and ◊i in order to obtain the truth lemma. The proof is the SMLKVr adaption of the proof of SELKVr in [13]. c Given a state s ∈ S such that ◊i φ ∈ s. We let Z = {ψ | ƒiψ ∈ s}∪{φ} and ƒc N X = {χ | i χ ∈ s}. Since X is countable, we list the elements in X as χi for i ∈ . c Note that since ⊢ ƒi ⊤, ⊤∈ X , namely X is non-empty. Fact 3.3 For any χ ∈ X , {χ}∪ Z is consistent. Therefore Z and every χ are also consistent.

Proof Suppose not, then there exists χ ∈ X , ψ1,..., ψn ∈ Z such that ⊢ ψ1 ∧···∧ NECK DISTK ψn ∧ φ →¬χ. By and , we have ⊢ ƒi(ψ1 ∧···∧ ψn) → ƒi(φ →¬χ). Kvr Since ƒiψ1,..., ƒi ψn ∈ s, ƒi(φ → ¬χ) ∈ s. Note that ∨ is equivalent to c c SUB c c ƒi(p → q) → (◊i p → ◊i q). By , we have ⊢ ƒi(φ →¬χ) ∧ ◊i φ → ◊i ¬χ. This c together with the fact that ƒi(φ →¬χ) ∈ s and ◊i φ ∈ s (assumption), we have c ◊i ¬χ ∈ s, contradiction. Since ⊤∈ X , {⊤}∪ Z is consistent thus Z is consistent.ƒ

Let B0 = Z ∪{χ0}, C0 = Z. We inductively construct Bn and Cn as following: • If Bn ∪{χn+1} is consistent, then Bn+1 = Bn ∪{χn+1}, Cn+1 = Cn. • Else, Bn+1 = Bn, Cn+1 = Cn ∪{χn+1}. • Finally, let B = Sn<ω Bn, C = Sn<ω Cn.

In order to show that B and C are consistent we first show that Bn and Cn are consistent for each n <ω. Before that we prove a handy proposition which will be useful soon. r Proposition 3.4 For any finitely many MLKv formulas ψ1,..., ψn, we have:

c c c ⊢ ƒi ψ1 ∧···∧ ƒi ψn ∧ ◊i(¬ψ1 ∧···∧¬ψn) → ƒi (ψ1 ∧···∧ ψn) 11

Kvr c c c Proof From ∨ we have ⊢ ƒi φ ∧ ƒi ψ ∧ ◊i(¬φ ∧¬ψ) → ƒi (φ ∧ ψ). What we want now is its generalization. By applying Kvr ∨, we can have the following sequence of SMLKVr theorems:

c c c ⊢ƒi ψ1 ∧ ƒi ψ2 ∧ ◊i(¬ψ1 ∧¬ψ2) → ƒi (ψ1 ∧ ψ2) c c c ⊢ƒi (ψ1 ∧ ψ2) ∧ ƒi ψ3 ∧ ◊i(¬(ψ1 ∧ ψ2) ∧¬ψ3) → ƒi (ψ1 ∧ ψ2 ∧ ψ3) · · · c c c ⊢ƒi (ψ1 ∧···∧ ψn−1) ∧ ƒi ψn ∧ ◊i(¬(ψ1 ∧···∧ ψn−1) ∧¬ψn) → ƒi (ψ1 ∧···∧ ψn)

Note that we have ⊢ ◊i(¬ψ1 ∧···∧¬ψn) → ◊i(¬(ψ1 ∧ ψk−1) ∧¬ψk) in standard normal modal logic, for any 2 ≤ k ≤ n. Now by using the above theorems one by one we can obtain:

c c c ⊢ ƒi ψ1 ∧···∧ ƒi ψn ∧ ◊i(¬ψ1 ∧···∧¬ψn) → ƒi (ψ1 ∧···∧ ψn) ƒ

Proposition 3.5 For any k ≥ 0, if Bk is consistent and χk+1 is not consistent with Bk, then χk+1 is consistent with Ck. Therefore Bk and Ck are consistent for k ∈ N.

Proof Suppose not, i.e., χk+1 is not consistent with both Bk and Ck. Let U = Bk \Z, V = Ck \ Z, U = {¬ψ : ψ ∈ U}, and V = {¬ψ : ψ ∈ V}. Then there exist α1,..., αℓ, β1,..., βm, γ1,..., γn ∈ Z such that: • ⊢ α1 ∧···∧ αℓ ∧ V U ∧ φ →¬χk+1 • ⊢ β1 ∧···∧ βm ∧ V V ∧ φ →¬χk+1 • ⊢ γ1 ∧···∧ γn ∧ V U ∧ φ → V V

The last one is due to the fact that any formula in Ck \ Z is inconsistent with Bk by construction. By NECK, DISTK and the definition of Z and X , we have

• ƒi(V U ∧ φ →¬χk+1) ∈ s • ƒi(V V ∧ φ →¬χk+1) ∈ s • ƒi(V U ∧ φ → V V ) ∈ s

First, we claim that ◊i(V U ∧ φ) ∈ s. If not, then ƒi¬(V U ∧ φ) ∈ s, which means that ¬(V U ∧ φ) ∈ Z ⊆ Bk. But as U ⊆ Bk, φ ∈ Bk, this implies that Bk is inconsistent, contradiction.

Then, we claim that ◊i(¬χk+1 ∧V V ) ∈ s. Since ƒi(V U ∧φ →¬χk+1) ∈ s and ƒi(V U ∧ φ → V V) ∈ s then ƒi(V U ∧ φ →¬χk+1 ∧ V V ), we immediately get ◊i(¬χk+1 ∧ V V ) ∈ s due to the fact ◊i(V U ∧ φ) ∈ s that we just showed. c c Finally, since ◊i(¬χk+1 ∧ V V ) ∈ s, ƒi χk+1 ∈ s and ƒi ψ ∈ s for all ψ ∈ V, by c Proposition 3.4, we have ƒi (χk+1 ∧V V) ∈ s. But since ƒi(V V ∧φ →¬χk+1) ∈ s, DISTKvr its equivalent ƒi(V V ∧χk+1 →¬φ) is also in s. By we know ⊢ ƒi(V V ∧ c c c χk+1 →¬φ) → ƒi (V V ∧ χk+1) → ƒi ¬φ. Therefore ƒi ¬φ ∈ s, contradiction to ◊c i φ ∈ s. Now, we can prove that Bk and Ck are consistent for any k ∈ N. We do induc- tion on k. For k = 0, then B0 = C0 = Z, whose consistency is shown in Fact 3.3. 12 “Knowing value” logic as a normal modal logic

For k = i + 1, consider whether χi+1 is consistent with Bi . If χi+1 is consistent with Bi , then Bk = Bi ∪ χi+1 and Ck = Ci (by induction hypothesis) are consistent. If χi+1 is inconsistent with Bi, then by induction hypothesis, Bi = Bi+1 and Ci are consistent. So according to the above conclusion Ci+1 is also consistent. ƒ

Proposition 3.6 B = Sn<ω Bn and C = Sn<ω Cn are both consistent.

Proof Suppose B is not consistent. That is, there exist φ1,..., φn ∈ B such that ⊢ φ1 ∧···∧φn →⊥. Therefore, there must be a finite m such that φ1,..., φn ∈ Bm. But this means that Bm is already inconsistent, contradictory to the construction of Bk. The case for C is similar. ƒ It is routine to prove the following:

Lemma 3.7 (Existence Lemma for ◊i) Given a state s ∈ S. If ◊iφ ∈ s, then there exists t ∈ S such that s →i t and φ ∈ t; c Also we have the existence lemma for ◊i : c c Lemma 3.8 (Existence Lemma for ◊i ) Given a state s ∈ S. If ◊i ψ ∈ s, then there c exist t, u ∈ S such that sRi tu and ψ ∈ t ∩ u. Proof Let Z, B and C be defined as above. Due to Proposition 3.6 B and C are both consistent. Therefore, both can be extended into maximal consistent sets, c say t and u. Now, the construction of B and C itself guarantee that sRi tu and φ ∈ t, u. ƒ

Lemma 3.9 (Truth Lemma) For any state s ∈M and φ, M , s  φ ⇐⇒ φ ∈ s.

c Proof Prove by induction. We only give the ◊i ψ case; the others are routine. ◊c c  ⇒: Suppose M , s i ψ. Then there exist t, u such that sRi tu, M , t ψ and c c M , u  ψ. By induction hypothesis, ψ ∈ t ∩ u. If ◊i ψ 6∈ s, then ƒi ¬ψ ∈ s, which c c implies ¬ψ ∈ t ∪ u by the construction of Ri , contradiction. Therefore, ◊i ψ ∈ s. c c ⇐:Suppose ◊i ψ ∈ s. Then according to the existence lemma for ◊i , there exist c t, u such that sRi tu and ψ ∈ t ∩ u. By induction hypothesis, M , t  ψ, M , u  ψ. c Therefore M , s ◊i ψ. ƒ The completeness result then follows immediately:

Theorem 3.10 (Completeness) SMLKVr is strongly complete over arbitrary mod- els.

Remark 3.11 At this point, it is interesting to compare our canonical model with the canonical model used in [13]. A complication in [13] is that merely max- imal consistent sets are not enough to build a FO canonical Kripke model. How- ever, as we have seen, we only use the maximal consistent sets in our canonical MLKvr model: it does not involve value assignments. Thus we have restored the symmetry between the logical language and the model to some extent: there is no longer too much information in the model, which cannot be talked about by the language. Note that we allow sRi tt, which also helps to have compact models. 13

4 Extended language with binary modalities c In the previous sections, we treat ◊i as a unary modality interpreted by a tern- c ary relation. Essentially, ◊i can be viewed as a binary modality where the two arguments are the same. In this section, we restore the symmetry between the ◊c semantics and the syntax one step further by having the binary i (·, ·) in the language. Surprisingly, this extension does not increase the expressive power of MLKvr . What is more, the new logic is normal. Consequently, the extension will help us to understand MLKvr more deeply from a normal modal logic point of view. The extended language MLKvb is given by the following BNF (b for binary):

c φ ::= ⊤ | p |¬φ | (φ ∧ φ) | ƒiφ | ƒi (φ, φ) ◊c ƒc ◊c b We define i (ψ, φ) as ¬ i (¬ψ, ¬φ). And i φ is now equivalent to the MLKv c c formula ◊i (φ, φ). To see the intuition, for example, ◊i (p, ¬p) says that i can see a p world and a ¬p world which do not agree on the value of c. Formally, the r c semantics is defined on the same MLKv models M = 〈S, {→i: i ∈ I}, {Ri : i ∈ I, c ∈ C}, V, VC〉:

c c M , s  ◊i (φ, ψ) ⇔ there exist t, u ∈ S such that sRi tu, M , t  φ and M , u  ψ.

The above semantics coincides with the standard semantics for binary diamond 8 c c modalities [1]. Note that ◊i (φ, ψ) is essentially different from ◊i (φ ∨ ψ): the latter only says that there are two φ ∨ψ-successors that have different values of c, but not necessarily one φ world and one ψ world. So, on first sight, MLKvr seems to be weaker than MLKvb. However, we will show by the following lemma that MLKvr and MLKvb are c c equally expressive, by reducing the binary ◊i to the unary ◊i in presence of the diamond ◊i . c Lemma 4.1 ◊i (φ, ψ) is equivalent to the disjunction of the following three formu- las: ◊c ◊ (i) i φ ∧ iψ c (ii) ◊i ψ ∧ ◊iφ ◊ ◊ ◊c ◊c ◊c (iii) iφ ∧ iψ ∧¬ i φ ∧¬ i ψ ∧ i (φ ∨ ψ) Proof The proof consists of two directions. ◊c First, we show that each of the three disjuncts entails i (φ, ψ). c (i) For any model M , s that satisfies ◊i φ ∧ ◊iψ, there exists t, u ∈ S such that c sRi tu, t  φ and u  φ, and exists v such that s →i v and v  ψ. According c c c to the property of Ri , at least one of sRi t v and sRi uv holds. W.l.o.g. suppose c c c sRi t v. Then according to the semantics of ◊i , we have s ◊i (φ, ψ).

8 Binary modalities appear in many modal logics, such as the until operator in temporal logic, and the relevant implication in relevance logic interpreted on Kripke models with a ternary relation. 14 “Knowing value” logic as a normal modal logic

c (ii) For ◊i ψ ∧ ◊iφ, the proof is similar to (i). c c c (iii) If M , s ◊i φ∧◊iψ∧¬◊i φ∧¬◊i ψ∧◊i (φ∨ψ), then: s has φ-successors and ψ-successors; all φ-successors have the same value of c, all φ-successors have c the same value of c, but the two values are different due to M , s ◊i (φ∨ψ). So we can easily guarantee that there are two states, one φ-successor and one ψ-successor of s such that they have different values with regard to c. This c means M , s ◊i (φ, ψ). c Second, we prove that if M , s ◊i (φ, ψ), then at least one of (i), (ii) and (iii) holds. c c Suppose M , s ◊i (φ, ψ), namely there exist t, u ∈M such that sRi tu, M , t  c φ and M , u  ψ. We immediately have M , s ◊iφ ∧ ◊iψ ∧ ◊i (φ ∨ ψ). If neither c c c c c ◊i φ nor ◊i ψ holds on s, then M , s  ◊iφ ∧ ◊iψ ∧¬◊i φ ∧¬◊i ψ ∧ ◊i (φ ∨ ψ). c c c c Therefore, M , s  (◊i φ ∧ ◊iψ) ∨ (◊i ψ ∧ ◊iφ) ∨ (◊iφ ∧ ◊iψ ∧¬◊i φ ∧¬◊i ψ ∧ c ◊i (φ ∨ ψ)) In sum, we can now conclude the equivalence. ƒ With this lemma in hand, the reduction theorem is straightforward: Theorem 4.2 (Reduction) For any MLKvb formula φ, there exists an MLKvr for- mula ψ such that for any pointed model M , s: M , s  φ ⇐⇒ M , s  ψ. Proof We define a reduction function r inductively:

• r(p)= p; r(¬φ)= ¬r(φ); r(φ ∧ ψ)= r(φ) ∧ r(ψ); r(◊iφ)= ◊i r(φ); • c c c r(◊i (φ, ψ)) = (◊i r(φ) ∧ ◊i r(ψ)) ∨ (◊i r(ψ) ∧ ◊i r(φ)) ∨ (◊i r(φ) ∧ ◊i r(ψ) ∧ ◊c ◊c ◊c ¬ i r(φ) ∧¬ i r(ψ) ∧ i (r(φ) ∨ r(ψ))). The correctness of the reduction is guaranteed based on Lemma 4.1. It is not hard (but important) to see that the rewriting always terminates. ƒ

r b Remark 4.3 Although MLKv is equally expressive as MLKv in presence of ◊i, it is not the case if ◊i is absent. To see this, consider the following two pointed c c c models M , s and N , x where sRi tu, sRi uv in the left model, and xRi yz in the other model: t : p y : p s9 r8 sis c rir sss rrr s ▲ i / u : p x ▲ c ▲▲▲ ▲▲▲ ▲ c ▲ i ▲▲% i ▲▲& v% : q z& : p

c We can use ◊i (p, q) to distinguish the two pointed models; however, they are c indistinguishable by using any formula with the unary ◊i but no ◊i, which can be proved by a simple induction. Now observe that MLKvb is a standard modal language defined on standard Kripke models with standard semantics. It is a relatively routine exercise to propose a normal modal logic system with the following axioms SYM, INCL and ATEUC to capture the corresponding special properties of the models: 15

Rules φ, φ → ψ System SMLKVb MP Axiom Schemas ψ φ TAUT all the instances of tautologies NECK ƒ DISTK ƒ ( → ) → (ƒ → ƒ ) iφ i p q i p iq φ DISTKvb ƒc(p → q, r) → (ƒc(p, r) → ƒc(q, r)) NECKvb i i i ƒc(φ, ψ) SYM ƒc(p, q) → ƒc(q, p) i i i φ INCL c SUB ◊i (p, q) → ◊i p ATEUC c c c φ[p/ψ] ◊i (p, q) ∧ ◊i r → ◊i (p, r) ∨ ◊i (q, r) RE ψ ↔ χ φ ↔ φ[ψ/χ] Note that due to SYM, we do not need to include the variations of DISTKvb NECKvb ƒc and w.r.t. the second argument in the binary i (cf. [1] for the standard proof systems of polyadic normal modal logics.) In this system SMLKVb we can derive all the axioms in SMLKVr . Before proving it, we first show the following handy propositions. c c c Proposition 4.4 ⊢SMLKVb ◊i (p ∨ q, r) → ◊i (p, r) ∨ ◊i (q, r). Proof This proposition captures the interaction between boolean operator ∨ and c 9 DISTKvb RE ◊i . So we can only start from the axiom . Note that is used fre- quently. c c c DISTKvb (1) ƒi (p → q, r) → (ƒi (p, r) → ƒi (q, r)) ( ) c c c (2) ¬(ƒi (p, r) → ƒi (q, r)) →¬ƒi (p → q, r) c c c (3) ƒi (p, r) ∧ ◊i (¬q, ¬r) → ◊i (p ∧¬q, ¬r) ◊c ƒc ◊c (4) i (¬q, ¬r) → (¬ i (p, r) ∨ i (p ∧¬q, ¬r)) c c c (5) ◊i (¬q, ¬r) → (◊i (¬p, ¬r) ∨ ◊i (p ∧¬q, ¬r)) c c c SUB (6) ◊i (p ∨ q, r) → (◊i (p, r) ∨ ◊i (¬p ∧ (p ∨ q), r)) ((5) & ) c c c (7) ◊i (p ∨ q, r) → (◊i (p, r) ∨ ◊i (q, r)) ƒ c c c Proposition 4.5 ⊢SMLKVb ◊i (p ∧ q, r) → ◊i (p, r) ∧ ◊i (q, r). Proof This is similar to the above proof. ƒ c c c Proposition 4.6 ⊢SMLKVb ƒi (p, r) ∧ ƒi (q, r) ∧ ◊i¬r → ƒi (p, q). ATEUC c c c Proof Easily derived from : ◊i (p, q) ∧ ◊i r → ◊i (p, r) ∨ ◊i (q, r). ƒ Proposition 4.7 All the SMLKVr axioms are provable in SMLKVb and the rules r b c c of SMLKV are admissible in SMLKV (viewing ◊i φ as ◊i (φ, φ)). Proof We need to check DISTKvr , Kvr ∨ and NECKvr in SMLKVr . (i) DISTKvr : ◊c ◊ INCL (1) i (p, q) → i p ( ) c (2) ƒi¬p → ƒi (¬p, ¬q) c SUB (3) ƒi(p → q) → ƒi (p → q, p) ((2) )

9 Actually this is a standard axiom for normal modal logic. In case the binary case might not be that familiar, we give the proof here. 16 “Knowing value” logic as a normal modal logic

c SUB (4) ƒi(p → q) → ƒi (p → q, q) ((2) ) c c (5) ƒi(p → q) → (ƒi (p → q, p) ∧ ƒi (p → q, q)) ((3) (4)) c c c c DISTKvb (6) ƒi(p → q) → (ƒi (p, p) → ƒi (q, p)) ∧ (ƒi (p, q) → ƒi (q, q)) ( ) c c SYM (7) ƒi(p → q) → (ƒi p → ƒi q) ((6) ) (ii) Kvr ∨: c c c ATEUC (1) ◊i (p, q) ∧ ◊i r → ◊i (q, r) ∨ ◊i (p, r) ( ) c c SUB (2) ◊i (p ∨ q, p ∨ q) ∧ ◊i(p ∧ q) → ◊i (p ∨ q, p ∧ q) ( ) c c c (3) ◊i (p ∨ q) ∧ ◊i(p ∧ q) → (◊i (p, p ∧ q) ∨ ◊i (q, p ∧ q)) (Prop. 4.4) c c c (4) ◊i (p ∨ q) ∧ ◊i(p ∧ q) → (◊i (p, p) ∨ ◊i (q, q)) (Prop. 4.5) c c c (5) ◊i (p ∨ q) ∧ ◊i(p ∧ q) → (◊i p ∨ ◊i q) (iii) NECKvr : It is a special case of NECKvb in SMLKVb where the two arguments are the same. Other axioms and rules in SMLKVr are exactly the same as in SMLKVb. ƒ Now as we can see below, the standard technique suffices to prove the com- pleteness of SMLKVb. The only tricky point is the ternary canonical relation.

Theorem 4.8 SMLKVb is sound and strongly complete w.r.t. MLKvr models. Proof The soundness is straightforward to check. For the completeness we build a canonical model:

c M = 〈S, {→i: i ∈ I}, {Ri : i ∈ I, c ∈ C}, VC〉

• S is the set of all maximal SMLKVb-consistent sets of MLKvb formulas, • s →i t ⇐⇒ {φ : ƒiφ ∈ s}⊆ t, • c c sRi tu ⇐⇒ (1) {φ : ƒiφ ∈ s} ⊆ t ∩ u and (2) for any ƒi (φ, ψ) ∈ s, φ ∈ t or ψ ∈ u. • VC(s)= {p : p ∈ s}. c Note that the existence lemma for ◊i is quite routine for normal polyadic c modal logic, cf. [1]. The idea is to build two i-successors t, u of s if ◊i (φ, ψ) ∈ s, c such that φ ∈ t, ψ ∈ u and sRi tu. According to the method in [1, pp. 200], we can build two maximal consistent sets t, u such that φ ∈ t and ψ ∈ u, and for all c c ƒi (χ1, χ2) ∈ s we have χ1 ∈ t or χ2 ∈ u. To make sure sRi tu we just need to INCL c check condition (1). To see this, note that by we have ƒiχ → ƒi (χ, θ) ∈ s SYM c and by we have ƒiχ → ƒi (θ, χ) ∈ s. Therefore for each ƒiχ ∈ s we have c c ƒi (χ, ⊥) ∈ s and ƒi (⊥, χ) ∈ s. Due to the construction of maximal consistent sets t, u, we have χ ∈ t or ⊥∈ u, and ⊥∈ t or χ ∈ u, which implies χ ∈ t ∩ u. Thus c {χ : ƒiχ ∈ s}⊆ t ∩ u. This concludes the proof that sRi tu. Based on the existence c lemmas for both ◊i and ◊i we can prove the truth lemma φ ∈ s ⇐⇒ s  φ using standard techniques. In the rest of this proof we verify that the canonical model satisfies the three r c properties of MLKv models. Note that condition (1) in the definition of Ri is symmetric, and condition (2) is also implicitly symmetric due to axiom SYM. Itis 17

c also obvious that sRi tu implies sRi t and sRi u by definition. We only need to verify the anti-euclidean property. c c c Towards contradiction suppose sRi tu, s →i v but neither sRi t v nor sRi uv. c c c Then according to the definition of Ri , there exist ƒi (φ1, ψ1), ƒi (φ2, ψ2) ∈ s such that ¬φ1 ∈ t, ¬ψ1 ∈ v, ¬φ2 ∈ u and ¬ψ2 ∈ v. Therefore ¬ψ1 ∧¬ψ2 ∈ v. DISTKvb SYM NECKvb Since s →i v, ◊i(¬ψ1 ∧¬ψ2) ∈ s. By , , and , it is not c c c hard to show ⊢SMLKVb ƒi (φ1, ψ1) → ƒi (φ1, ψ1 ∨ ψ2) and ⊢SMLKVb ƒi (φ2, ψ2) → c c c ƒi (φ2, ψ1 ∨ψ2). So ƒi (φ1, ψ1 ∨ψ2), ƒi (φ2, ψ1 ∨ψ2) ∈ s. By Proposition 4.6 and SUB c c c , ⊢SMLKVb ƒi (φ1, ψ1∨ψ2)∧ƒi (φ2, ψ1∨ψ2)∧◊i (ψ1∨ψ2) → ƒi (φ1, φ2). Since c c c ƒi (φ1, ψ1 ∨ψ2), ƒi (φ2, ψ1 ∨ψ2) and ◊i(ψ1 ∨ψ2) are all in s, ƒi (φ1, φ2) ∈ s. This c together with sRi tu imply that φ1 ∈ t or φ2 ∈ u, contradictory to the assumption that ¬φ1 ∈ t and ¬φ2 ∈ u. ƒ

5 Applications 5.1 Bisimulation In the field of modal logic, various bisimulation notions help to characterize the expressive power of the new semantics-driven logics. As a normal modal logic, MLKvb has a natural notion of bisimulation (cf. [1]), and it will in turn help us to find a notion of bisimulation over FO Kripke models for the original ELKvr .

1 c Definition 5.1 (C-Bisimulation) Let M1 = 〈S1, {→i : i ∈ I}, {Ri : i ∈ I, c ∈ 2 c b C}, V1〉, M2 = 〈S2, {→i : i ∈ I, c ∈ C}, {Qi : i ∈ I}, V2〉 be two models for MLKv r (also for MLKv ). A C-bisimulation between M1 and M2 is a non-empty binary Z Z relation ⊆ S1 × S2 such that for all s1 s2, the following conditions are satisfied:

Inv : V1(s1)= V2(s2); 1 2 Z Zig : s1 →i t1 ⇒∃t2 such that s2 →i t2 and t1 t2; 2 1 Z Zag : s2 →i t2 ⇒∃t1 such that s1 →i t1 and t1 t2; c Z Z c Kvb-Zig : s1Ri t1u1 ⇒∃t2, u2 ∈ S2 such that t1 t2, u1 u2 and s2Qi t2u2; c Z Z c Kvb-Zag : s2Qi t2u2 ⇒∃t1, u1 ∈ S1 such that t1 t2, u1 u2 and s1Ri t1u1.

We say M , s and N , t are C-bisimilar (M , s↔C N , t) if there is a C-bisimulation Z between M and N and (s, t) ∈ Z. ↔ Theorem 5.2 If M1, s1 C M2, s2, then M1, s1 ≡MLKvb M2, s2.

Proof Suppose M1, s1 ↔C M2, s2. We prove by induction on the structure of b c MLKv formulas, and the only non-trivial case is when φ = ◊i (ψ, χ). c c Suppose M1, s1 ◊i (ψ, χ), then ∃t1, u1 ∈ S1 such that s1Ri t1u1 with M1, t1  Z Z ψ and M1, u1  χ. By Kvb-Zig, there exist t2, u2 ∈ S2 such that t1 t2, u1 u2 c and s2Qi t2u2. By induction hypothesis, M2, t2  ψ and M2, u2  χ. Therefore, c M2, s2 ◊i (ψ, χ). The other side is similar by Kvb-Zag. ƒ As in normal modal logic, we have the following theorem for MLKvb (we omit c the rather standard proof, but one can try to see how the binary modality ◊i facilitates the proof): 18 “Knowing value” logic as a normal modal logic

Theorem 5.3 Suppose M , N are finite models. Then M , s ↔C N , t ⇐⇒ M , s ≡MLKvb N , t. Since MLKvr and MLKvb have the same expressive power we immediately have:

Corollary 5.4 Suppose M and N are finite models. Then M , s ↔C N , t ⇐⇒ M , s ≡MLKvr N , t. In [12], a notion of bisimulation has been offered for ELKv, the epistemic logic r with unconditional Kvi operators. However, for ELKv it was not that clear about the suitable bisimulation notion. Now we can recast C-bisimulation back to the setting of ELKvr over FO Kripke models since ELKvr and MLKvr are essentially the same language. Definition 5.5 (C-bisimulation over FO Kripke models) Given two pointed FO 1 1 2 Kripke models M = 〈S1, D1, {→i : i ∈ I}, V1, VC 〉, and N = 〈S2, D2, {→i : i ∈ 2 Z I}, V2, VC 〉, a relation ⊆ S1 × S2 is a C-bisimulation between the two models M , N Z if whenever s1 s2 we have:

Inv V1(s1)= V2(s2); 1 2 Z Zig : s1 →i t1 ⇒∃t2 such that s2 →i t2 and t1 t2; 2 1 Z Zag : s2 →i t2 ⇒∃t1 such that s1 →i t1 and t1 t2; 1 1 1 1 Kvr-Zig If s1 →i t1 and s1 →i u1 and VC (c, t1) 6= VC (c, u1) then there are t2 and 2 2 2 2 u2 in N such that s2 →i t2, s2 →i u2, t1 Zt2, u1 Zu2, and VC (c, t2) 6= VC (c, u2) in N . 2 2 2 2 Kvr-Zag If s2 →i t2 and s →i u2 and VC (c, t2) 6= VC (c, u2) then there are t1 and u1 1 1 1 1 in M such that s1 →i t1, s →i u1, t1Zt2, u1 Zu2 and VC (c, t1) 6= VC (c, u1) in M .

Abusing the notation, FO Kripke models M , s and N , t are C-bisimilar (M , s ↔C N , t) iff there exists a C-bisimulation Z between M and N such that s, t ∈ Z. Now since MLKvb and MLKvr have exactly the same expressive power, and MLKvr is equivalent to ELKvr modulo translation. The above C-bisimilation works for ELKvr , as proved in detailed in [6]:

Theorem 5.6 For finite FO Kripke models M1, M2: M1, s1 ↔C M2, s2 iff M1, s1 ≡ELKvr M2, s2. 5.2 Completeness of SMLKV The unconditional Kv operator was introduced in [7] in the context of epistemic logic (call the language ELKv):

φ ::= ⊤ | p |¬φ | (φ ∧ φ) | Kiφ | Kvi c r r Essentially, Kvi c is Kvi (⊤, c) in ELKv . The semantics is as in the case of ELKv , which is based on FO Kripke models. Plaza gave two axioms on top of S5 which are the counterparts of the introspection axioms in standard epistemic logic (over FO epistemic models):

Kvi c → KiKvi c ¬Kvi c → Ki¬Kvi c 19

However, neither [7] nor [12,13] gave a complete proof of this simple logic. Here c we look at this language from our ◊i perspective, and consider the corresponding simple language (MLKv) over the class of all the models:

c φ ::= ⊤ | p |¬φ | (φ ∧ φ) | ƒiφ | ƒi ⊥

c c Note that ¬Kvi c can be viewed as ◊i ⊤. Thus Kvi c is indeed ƒi ⊥. r c The semantics is just as in MLKv but we only allow ⊤ as the argument for ◊i . As in the case of ELKvr and MLKvr we can simply show that: Proposition 5.7 For any set of ELKv formula Σ ∪{φ}, Σ  φ iff T(Σ)  T(φ). Now based on this view, a natural system SMLKV is obtained by simplifying the system SMLKVr :

Rules → MP φ, φ ψ System SMLKV ψ Axiom Schemas NECK φ TAUT all the instances of tautologies ƒiφ DISTK ƒ ƒ ƒ φ i(p → q) → ( i p → i q) SUB INCLT c ◊i ⊤→ ◊i ⊤ φ[p/ψ] ↔ RE ψ χ φ ↔ φ[ψ/χ] ◊c ◊c ƒc Note that due to the fact that the only i formula is i ⊤ (and i ⊥), most of the previous axioms and rules do not apply. We only need to add one axiom INCLT on top of the usual normal modal logic, inspired by the INCL axioms of SMLKVb. We go on to prove the completeness of SMLKV. c Definition 5.8 The canonical model M is a tuple 〈S, {→i : i ∈ I}, {Ri : i ∈ I, c ∈ C}, V 〉 where: • S is the set of maximal SMLKV-consistent sets, • s →i t ⇐⇒ {φ : ƒiφ ∈ s}⊆ t, • c ′ ′ c sRi tt ⇐⇒ s →i t, s →i t and ◊i ⊤∈ s, • V(s)= {p : p ∈ s} c The only tricky point is the definition of canonical relations Ri . The intuition is that as long as a state can see two states having different values, then we can safely assume all the states that it can see have different values. Note that in r c r MLKv models we also allow sRi tt. We first need to verify that M is an MLKv model: Proposition 5.9 The canonical model M is an MLKvr model. Proof We only need to verify the three conditions. The first two are again obvious c ′ by definition, so we only prove the anti-euclidean property. Suppose sRi tt and ′ c c c ′ s →i u. Then s →i t, s →i t and ◊i ⊤ ∈ s. So both sRi tu and sRi t u by the c definition of Ri . ƒ 20 “Knowing value” logic as a normal modal logic

c The existence lemma for ◊i is routine. As for the case of ◊i : c c Lemma 5.10 (Existence Lemma for ◊i ) If ◊i ⊤ ∈ s, then there exist t, u such that c sRi tu. c INCLT Proof Suppose ◊i ⊤∈ s then due to ◊i⊤∈ s. By the existence lemma for c ◊i, it follows that there exists t such that s →i t. Therefore by definition sRi tt. ƒ Lemma 5.11 For any s in M and MLKv formula φ, M , s  φ ⇐⇒ φ ∈ s. c Proof The only interesting case is when φ = ◊i ⊤. c ′ c ′ ⇒: Suppose M , s  ◊i ⊤. Then there exist t, t such that sRi tt . By the c c definition of Ri , ◊i ⊤∈ s. c c ′ ⇐: Suppose ◊i ⊤∈ s. By the existence lemma for ◊i , there exist t, t such that c ′ c sRi tt . Therefore M , s ◊i ⊤. ƒ Theorem 5.12 SMLKV is strongly complete w.r.t. MLKv models. A corollary follows immediately based on Proposition 5.7: INCLT ˆ Corollary 5.13 SMLKV (viewing as ¬Kvi⊤ → Ki⊤) is strongly complete w.r.t. ELKvr models.

6 Discussion and future work In this paper, we introduce a ternary relation based simple semantics to the “know- ing value” logic without explicit value assignments. Under this semantics, the logic can be viewed as a disguised normal modal logic with both standard unary and binary modalities. The use of this perspective is demonstrated by various applica- tions. Another intuitive way to simplify the original FO-Kripke semantics is to in- troduce a binary relation ≍c for each c representing the inequality of the value of c. Correspondingly, in the language, besides ◊iφ we may introduce ◊cφ for- mulas saying that there is a different world where φ holds but c has a different value compared to the current world. However, it is not straightforward to express

¬Kv(ψ, c) in this language. The closest counterpart ◊i(ψ ∧ ◊cψ) will not do the job alone. We probably need to add a further condition: ◊i◊c p → ◊i p which says the ≍c successors of an i-reachable world are again i-reachable. Actually it means c that we should combine →i and ≍c which is almost our ternary Ri . Moreover, to axiomatize this ≍c we need the axioms of anti-equivalence (irreflexivity, sym- metry, and anti-euclidean property 10 ). However, irreflexivity and anti-euclidean property for the binary ≍c are not definable in modal logic. We probably need to ◊c Kv do the same as in the i case: use ∨ to capture the i-accessible anti-euclidean property to some extent. Having said the above, it is clear that our approach in this paper is more intuitive and technically natural. To close, we list a few directions which we leave for future occasions:

10 Here anti-euclidean property means if x ≍c y and there is another world z then x ≍c z or y ≍c z. A simple disjoint union argument can show it is not modally definable. Thanks to Zhiguang Zhao for pointing it out. 21

• The corresponding results in the setting of epistemic (S5) models. • Characterization theorem of ELKvr (MLKvr ) within first-order modal logic via C-bisimulation. • A decision procedure for ELKvr (MLKvr ) based on the simplified models. • In similar ways, we can try to simplify the semantics for other “knowing-X” logics, such as knowing whether, knowing how, and so on.

References

[1] Blackburn, P., M. de Rijke and Y. Venema, “Modal Logic,” Cambridge University Press, 2002. [2] Ding, Y., Axiomatization and complexity of modal logic with knowing-what operator on model class K (2015), unpublished manuscript. URL http://www.voidprove.com/research.html [3] Fagin, R., J. Halpern, Y. Moses and M. Vardi, “Reasoning about knowledge,” MIT Press, Cambridge, MA, USA, 1995. [4] Fan, J., Y. Wang and H. van Ditmarsch, Almost necessary, in: Advances in Modal Logic Vol. 10, 2014, pp. 178–196. [5] Fan, J., Y. Wang and H. van Ditmarsch, Contingency and knowing whether, The Review of Symbolic Logic 8 (2015), pp. 75–107. [6] Gattinger, M., J. van Eijck and Y. Wang, Knowing value and public inspection (2016), manuscript. [7] Plaza, J. A., Logics of public communications, in: M. L. Emrich, M. S. Pfeifer, M. Hadzikadic and Z. W. Ras, editors, Proceedings of the 4th International Symposium on Methodologies for Intelligent Systems, 1989, pp. 201–216. [8] van der Hoek, W.and A. Lomuscio, A logic for ignorance, Electronic Notes in Theoretical Computer Science 85 (2004), pp. 117–133. [9] van Ditmarsch, H., J. Halpern, W. van der Hoek and B. Kooi, editors, “Handbook of Epistemic Logic,” College Publications, 2015. [10] Wang, Y., A logic of knowing how, in: Proceedings of LORI 2015, 2015, pp. 392–405. [11] Wang, Y., Beyond knowing that: a new generation of epistemic logics, in: Hintikka’s volume in outstanding contributions to logic, Springer, 2016 Forthcoming. URL http://arxiv.org/abs/1605.01995 [12] Wang, Y. and J. Fan, Knowing that, knowing what, and public communication: Public announcement logic with Kv operators, in: Proceedings of IJCAI 13, 2013, pp. 1139–1146. [13] Wang, Y. and J. Fan, Conditionally knowing what, in: Advances in Modal Logic Vol. 10, 2014, pp. 569–587.